A negative pressure circulating explosion-proof dust removal method suitable for combustible dust sites

Abstract

The present application relates to the field of safe and environmentally friendly technology, and more particularly to a negative pressure circulation explosion prevention and dust removal method suitable for combustible dust sites, which comprises: first, sealing the warehouse body and injecting carbon dioxide-nitrogen inert gas to generate a low-oxygen environment; then driving the adjustable guide angle cyclone separator, dielectric barrier discharge plasma reactor and waste heat thermoacoustic transducer to cooperate in dust capture, and using a microphone to compressively sense the amplitude of the sound node to synthesize the capture data; then inputting the environmental data and the capture data into a reinforcement learning process containing a differentiable thermochemical model to output instructions for adjusting the negative pressure pulse, inert flow, sound field intensity and cyclone geometry; in combination with distributed radio frequency imaging data, an adversarial digital twin model is generated to predict the explosion free energy and automatically rollback or shut down when the threshold is exceeded, forming a safe, efficient and low-energy consumption closed loop.

Description

Technical Field

[0001] This invention relates to the field of safety and environmental protection technology, and in particular to a negative pressure circulation explosion-controlled dust removal method suitable for combustible dust locations. Background Technology

[0002] In industries involving the processing of metal powders such as aluminum, magnesium, and titanium, as well as lightweight combustible dusts such as pharmaceutical powders and wood flour, dust collection systems must not only ensure environmental emissions meet standards but also suppress the risk of dust cloud explosions. Traditional bag filters or positive pressure pulse dust collectors rely heavily on high airflow to transport dust to the filter bags, making it impossible to maintain a consistently low-oxygen environment. Once the filter material becomes clogged or the fan fluctuates, the coupling of local dust concentration with an ignition source can easily induce deflagration, causing equipment explosions, production line shutdowns, and personnel injuries, placing enormous pressure on industry safety and green production. Summary of the Invention

[0003] To address the numerous problems existing in the prior art, this invention provides a negative pressure circulating explosion control and dust removal method suitable for combustible dust environments. This invention constructs a state vector from multi-source sensor data, generates coupled commands through a differential thermochemical-reinforcement learning joint control module, and simultaneously utilizes radio frequency three-dimensional imaging to drive digital twin prediction of free energy and trigger rapid rollback or shutdown, thereby achieving simultaneous optimization of the three objectives of dust removal, explosion suppression, and energy consumption.

[0004] A negative pressure circulating explosion-controlled dust removal method suitable for combustible dust environments includes the following steps:

[0005] The sealed dust removal system creates negative pressure by suction and injects an inert mixture of carbon dioxide and nitrogen to lower the oxygen volume fraction below the safe threshold. It then collects negative pressure value, oxygen volume fraction, and temperature-strain information to generate environmental data.

[0006] Under the low-oxygen conditions, the adjustable guide angle cyclone separator, the dielectric barrier discharge plasma reactor, and the waste heat driven thermoacoustic transducer are operated in synergy. The microphone array measures the sound field and the compressed sensing method is used to reconstruct the sound node amplitude. The sound node amplitude is combined with the outlet dust concentration and the combustible gas concentration to form the capture data.

[0007] The environmental data and the captured data are combined to form a state vector, which is then input into a control flow containing a differential thermochemical model and a reinforcement learning network to generate and issue control commands for adjusting negative pressure pulses, inert gas flow rates, sound field intensity, and cyclone geometry parameters.

[0008] Based on the environmental data, the control commands, and the radio frequency imaging data acquired by the distributed antenna array, the explosion free energy is predicted using a generative adversarial digital twin model. When the current or predicted explosion free energy exceeds a threshold, a strategy rollback command or an emergency shutdown command is sent to achieve closed-loop control.

[0009] Preferably, the negative pressure value is obtained by an absolute differential pressure sensor, the oxygen volume fraction is obtained by a zirconium-based oxygen sensor, and the temperature-strain information is obtained by fiber Bragg gratings arranged in segments along the wall of the dust removal system, and the data are collected with a unified timestamp to form environmental data.

[0010] Preferably, before injection, the inert gas mixture is continuously mixed with carbon dioxide and nitrogen at a preset volume fraction by a mass flow control device, and the mixing flow rate is adjusted in real time according to the oxygen volume fraction to keep the oxygen volume fraction below a safe threshold.

[0011] Preferably, the microphone array is arranged at a non-uniform angle along the outer wall of the adjustable guide angle cyclone separator, and the sound pressure data acquired by the microphone array is sparsely reconstructed using the compressed sensing method to obtain the sound node amplitude.

[0012] Preferably, the adjustable guide angle cyclone separator synchronously adjusts the guide angle and cone ratio through an electric actuator. The guide angle and cone ratio are adjusted in a closed loop according to the real-time changes in the acoustic node amplitude and the outlet dust concentration.

[0013] Preferably, the dielectric barrier discharge plasma reactor adjusts the voltage between electrodes and the electrode spacing to match the discharge intensity with the oxygen volume fraction, thereby maintaining a stable ionization state and suppressing spark generation.

[0014] Preferably, the waste heat-driven thermoacoustic transducer receives waste heat generated by the dielectric barrier discharge plasma reactor and releases acoustic energy at a fixed acoustic frequency to enhance the coagulation of fine dust particles.

[0015] Preferably, a microthermochemical model can be used to calculate the explosion free energy and system entropy and output a gradient. A reinforcement learning network updates the control strategy based on the gradient to generate control commands for adjusting the negative pressure pulse, inert gas flow rate, sound field intensity, and cyclone geometry parameters.

[0016] Preferably, the reinforcement learning network includes a graph neural network for describing the coupling relationships between devices within the dust removal system, and a time-series encoder for processing the time-series features of environmental data and captured data.

[0017] Preferably, the distributed antenna array is deployed in the form of a planar array. It obtains radio frequency reflection amplitude and phase information by scanning the working area through phase modulation to form radio frequency imaging data. It generates an adversarial digital twin model, combines fluid dynamics calculation and discrete element calculation, and perturbs the Green tensor to simulate the most unfavorable working condition. When the current or predicted explosion free energy exceeds the threshold, the emergency shutdown command raises the negative pressure pulse and inert gas flow rate to the preset maximum safe value, and keeps the sound field intensity and the guide angle and cone ratio of the adjustable guide angle cyclone separator at the most recent safe state.

[0018] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:

[0019] By using sealed negative pressure and carbon dioxide-nitrogen inerting, a real-time low oxygen line was achieved, eliminating the oxygen supply chain; by coordinating an adjustable flow angle cyclone separator, plasma reactor, and thermoacoustic transducer, a three-field joint capture of acoustic, electrical, and centrifugal fields was achieved, significantly improving the efficiency of submicron dust removal; by using graph neural network-temporal coding reinforcement learning, combined with a differential thermochemical model, the optimal scheduling of full coupling of negative pressure, inerting, acoustic field, and geometry was achieved, reducing the peak free energy and saving energy; by using distributed radio frequency imaging and generative adversarial digital twin prediction, sub-second free energy over-threshold early warning and two-stage shutdown were achieved, maintaining closed-loop safety even under extreme disturbances. Attached Figure Description

[0020] Figure 1 This is a schematic flowchart of the method of the present invention;

[0021] Figure 2 This is a schematic diagram of the cooperative trapping unit in this invention;

[0022] Figure 3 This is a schematic diagram of the state vector and control closed loop in this invention. Detailed Implementation

[0023] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation.

[0024] like Figure 1 As shown, a negative pressure circulating explosion-controlled dust removal method suitable for combustible dust environments includes the following steps:

[0025] The sealed dust removal system creates negative pressure by suction and injects an inert mixture of carbon dioxide and nitrogen to lower the oxygen volume fraction below the safe threshold. It then collects negative pressure value, oxygen volume fraction, and temperature-strain information to generate environmental data.

[0026] After sealing the dust removal system, establishing a relatively independent controlled space to prevent continuous infiltration of external air is the prerequisite for achieving oxygen and pressure control in this invention. The suction device uses a combination of multi-stage impellers and venturi tubes to create continuous negative pressure, making the absolute pressure inside the chamber lower than the external atmospheric pressure. The negative pressure mechanism can be understood from two aspects: First, negative pressure tends to draw in rather than expel external leakage. Even if micro-cracks appear in the local seal due to vibration, it is more likely to draw in clean air rather than eject dust-laden gas, reducing dust dispersion. Second, negative pressure reduces the partial pressure of oxygen inside the chamber. Under the condition that the oxygen content remains unchanged in the same volume, the low-pressure environment increases the lower limit of dust combustion, leaving a safe margin for dilution by inert gases.

[0027] The inert gas mixture consists of carbon dioxide and nitrogen, achieving a balance between oxygen scavenging, cooling, and economy. Carbon dioxide has a high volumetric heat capacity and low thermal conductivity, enabling it to rapidly absorb heat and suppress temperature at localized hot spots; nitrogen has low density and low viscosity, facilitating rapid and uniform diffusion. The two gases are mixed and continuously injected into a mass flow control device according to a preset volume fraction, with the real-time flow rate adjusted by a closed-loop control based on the oxygen volume fraction. The dilution process can be approximated as follows:

[0028]

[0029] in Indicates time oxygen volume fraction, For inert gas flow rate, For the volume of the warehouse, The target oxygen volume fraction is given by this formula. This formula assumes that the inert gas is fully dispersed within a single feedback cycle and is applicable to scenarios with a storage capacity of 10 to 100 cubic meters and an inert gas flow rate of 100 liters per minute. By comparing the predicted value with the actual sensor measurement, the controller can automatically correct the flow rate in each feedback cycle to ensure that the oxygen volume fraction remains below the safe threshold.

[0030] The key sensor system consists of an absolute differential pressure sensor, a zirconium-based oxygen sensor, and a fiber Bragg grating. The absolute differential pressure sensor is installed on the top of the silo and connected to an external reference pipeline, providing the real-time difference between the static pressure inside the silo and the ambient pressure. The zirconium-based oxygen sensor utilizes the conductivity of a solid electrolyte for oxygen ions, calculating the oxygen volume fraction by measuring the potential generated by the oxygen partial pressure difference across the silo. The fiber Bragg gratings are arranged along the wall in a spiral or serpentine pattern. The reflected wavelength of each grating varies with temperature and strain; temperature and structural strain curves are obtained through spectral demodulation. The temperature and strain data can provide early warnings of localized slow combustion in the dust layer or thermal expansion of the metal casing, aiding in the assessment of seal integrity.

[0031] To ensure accurate reconstruction of operating conditions at the same moment in subsequent steps, this invention employs a unified timestamp strategy: the controller is configured with a high-precision real-time clock, sampling by each sensor is simultaneously initiated by a unified trigger pulse, each set of data is accompanied by a nanosecond-level timestamp and then enters the edge buffer before being pushed to the host computer via a message queue. In this way, environmental data remains strictly synchronized during multi-node transmission, avoiding operating condition mismatches caused by network latency.

[0032] This step produces three technical effects. First, the dual explosion suppression mechanism of sealed negative pressure and low oxygen simultaneously constrains the oxygen supply and dust suspension state of the three elements of combustion, significantly reducing the probability of spontaneous combustion. Second, the inert gas mixture provides a heat capacity buffer while maintaining low oxygen levels, and has an endothermic cooling effect on short-term high-temperature spots, which can suppress the ignition of fire sources. Finally, the temperature-strain distribution and pressure-oxygen data together form environmental data, providing highly reliable input for subsequent acoustic coagulation, plasma charging, and thermoacoustic enhancement, ensuring a coherent link from physical suppression to intelligent control.

[0033] Example 1: In a magnesium powder workshop renovation project, the effective volume of the silo was 20 cubic meters, and the negative pressure was set at 2000 Pa. The mass flow control device was configured according to an oxygen volume fraction threshold of 14%. When the oxygen volume fraction rose to 13%, the flow rate of the inert gas mixture was automatically increased. When the oxygen volume fraction dropped to 12% and remained there for 300 seconds, the system entered a pressure-maintaining and oxygen-maintaining mode. During the six-month operation period, no oxygen volume fraction exceeding the limit event occurred, and the dust emission was found to be below the industry limit by a third-party test. This demonstrates that the first step of this invention not only provides a stable and safe boundary but also provides a reliable environmental foundation for subsequent collaborative capture and digital twin prediction.

[0034] Preferably, the negative pressure value is obtained by an absolute differential pressure sensor, the oxygen volume fraction is obtained by a zirconium-based oxygen sensor, and the temperature-strain information is obtained by fiber Bragg gratings arranged in segments along the wall of the dust removal system, and the data are collected with a unified timestamp to form environmental data.

[0035] Real-time control of negative pressure circulating explosion-controlled dust removal relies on continuous sensing of system pressure, oxygen content, and structural thermo-mechanical state. This invention uses an absolute differential pressure sensor, a zirconium-based oxygen sensor, and a fiber Bragg grating to form a three-dimensional monitoring network. This network uniformly encodes the three types of raw measurements and writes them into "environmental data," providing high-confidence input for subsequent acoustic-plasma-thermoacoustic coordinated capture and digital twin prediction. The following section, in conjunction with the application scenarios of this invention, explains the principles, integration methods, and technical effects of the three sensors.

[0036] When the absolute differential pressure sensor is working, the static pressure inside the chamber and the atmospheric pressure in the reference chamber are simultaneously applied to the silicon micromechanical diaphragm. The change in diaphragm capacitance is converted into a differential pressure value using a Wheatstone bridge. In continuous suction mode, if the seal is intact, the differential pressure should remain stable within the set negative pressure range. Micro-leakage or pipe blockage can cause the differential pressure to drift slowly. The algorithm can trigger maintenance alerts at an early stage by judging the first derivative of the differential pressure and the average fluctuation threshold.

[0037] The zirconium-based oxygen sensor is a solid electrolyte oxygen analyzer, with its core tube exposed to the measuring gas and reference gas environments on opposite sides. When the temperature exceeds a critical value, oxygen ions migrate from the high partial pressure side to the low partial pressure side within the zirconium ceramic, creating a potential difference between the platinum electrodes due to the migrating charges. According to the Nernst equation:

[0038]

[0039] The partial pressure of oxygen on the measuring side can be calculated. For potential difference, The gas constant is Absolute temperature It is Faraday's constant. and These represent the oxygen partial pressures on the reference and measurement sides, respectively. The circuit... After sampling, the oxygen volume fraction can be obtained by referring to a table. Compared with traditional paramagnetic or electrochemical consumable probes, zirconium-based probes have a faster response speed and no electrolyte loss under micro-negative pressure and dusty conditions, making them suitable for long-term operation.

[0040] A fiber Bragg grating is constructed by writing periodic refractive index perturbations onto an ultraviolet phase mask, forming a wavelength-selective reflection structure. The center wavelength of the grating reflection satisfies the following condition:

[0041]

[0042] in For the Bragg wavelength, For effective refractive index, This refers to the grating period. When the wall surface containing the grating is heated or subjected to pressure, resulting in strain... and Synchronous changes, making Linear drift. The spectrometer scans all gratings at a fixed repetition frequency and outputs temperature-strain curves. Compared to point thermocouples, distributed gratings can achieve centimeter-level spatial resolution and provide earlier observability for hot spots caused by local dust accumulation; compared to resistance strain gauges, optical fibers are insensitive to electromagnetic interference and can still operate stably in plasma electric fields.

[0043] This invention employs a unified timestamp aggregation principle. The controller has a built-in pulse synchronization clock that drives the three analog-to-digital conversion channels via interrupts, assigning values ​​to all data frames after a single acquisition is completed. To compensate for the slight time difference caused by hardware delays in each channel, the absolute differential pressure channel delay is pre-calibrated. Oxygen channel delay Fiber optic channel delay The system performs the following steps on the original timestamp:

[0044]

[0045] Correction ( Pick , , The corrected values ​​are then written into the data frame header. This ensures that the generated environmental data is delivered to the host computer with nanosecond-level synchronization error. This time consistency eliminates the need for interpolation compensation in subsequent control processes when calculating state vectors, simplifying the algorithm and avoiding misjudgments caused by mismatches.

[0046] Example 2: Forty-eight fiber Bragg gratings are deployed on a 30-cubic-meter aluminum powder dust collection chamber, with a sampling frequency of 100Hz; the sampling frequency for oxygen content and differential pressure channels is 10Hz. When fluctuations in the feed rate cause a momentary drop in negative pressure, the controller detects this within two sampling cycles. The system lowers its speed and quickly adjusts the fan frequency, but the oxygen volume fraction does not increase significantly. If the moisture content of the dust in the same batch is too high, causing localized clogging of the filter bags, the fiber optic network detects a synchronous increase in temperature and strain at six adjacent measuring points, at which point the system can enter the pre-rinsing process. Actual test results show that, compared to the control group without Bragg grating monitoring, the use of this invention's integrated sensor solution reduces the number of erroneous shutdowns of the dust removal facility by 40% and extends the filter bag lifespan by 1.3 times.

[0047] Preferably, before injection, the inert gas mixture is continuously mixed with carbon dioxide and nitrogen at a preset volume fraction by a mass flow control device, and the mixing flow rate is adjusted in real time according to the oxygen volume fraction to keep the oxygen volume fraction below a safe threshold.

[0048] The injection control of the inert gas mixture is a key step in establishing a stable low-oxygen environment in this invention. The binary combination of carbon dioxide and nitrogen can simultaneously meet the comprehensive requirements of diluting oxygen concentration, endothermic temperature suppression, and economic feasibility. The mass flow control device adopts an integrated structure of a two-channel thermal mass flow meter and a proportional valve, with one inlet connected to a carbon dioxide cylinder and the other to a membrane-separated nitrogen manifold. The internal microprocessor receives real-time data from the oxygen sensor once per second, compares the oxygen volume fraction with a preset safety threshold, and calculates the mixed flow correction. This process is a closed-loop feedback control, and this invention introduces first-order predictive compensation, which retains the steady-state accuracy of proportional-integral-derivative control while improving the response speed to sudden load fluctuations.

[0049] Let the internal volume of the warehouse be Set the oxygen volume fraction threshold to Real-time oxygen volume fraction is Let the instantaneous flow rates of carbon dioxide and nitrogen be respectively... and The total flow rate of the inert gas mixture is According to the law of conservation of dilution, the rate of change of oxygen concentration can be obtained as follows:

[0050]

[0051] In the formula This represents the residual oxygen volume fraction in the gas mixture, theoretically close to zero. This expression reflects that the rate of decrease in oxygen volume fraction is proportional to the inert gas flow rate. The mass flow control device discretizes this equation and displays the change in oxygen volume fraction from the previous time step. As a feedforward term, the adjusted flow command is obtained:

[0052]

[0053] The proportionality coefficient Integral coefficient Differential coefficients From experimental tuning, feedforward coefficients The flow rate is automatically calculated based on the chamber volume and ventilation rate. The above formula only provides the flow rate for a single channel; in actual control, it is based on the volume ratio of carbon dioxide to nitrogen. Distribute the carbon dioxide flow rate as follows: Nitrogen flow rate is .

[0054] Compared to traditional manual valves or simple proportional valve opening control, the continuous mass flow control device of this invention has four significant advantages. First, the thermal mass flow meter directly measures the thermal conduction effect of gas molecules, making it insensitive to pressure fluctuations and providing higher flow accuracy than volumetric or differential pressure methods. Second, carbon dioxide and nitrogen are metered separately before the regulating valve, ensuring a stable volume fraction ratio and avoiding concentration shifts caused by differences in molecular weight after mixing. Third, the feedforward term... It can increase the flow rate and shorten the transition time in the early stages of an increase in oxygen volume fraction. Fourth, the microprocessor has a built-in linearization correction curve for the flow meter, which can automatically compensate for measurement drift caused by temperature changes.

[0055] When applying this invention, there is a coupling between the negative pressure suction rate of the dust removal system and the injection volume of the inert gas mixture. If the suction rate increases while the inert gas flow rate remains constant, the oxygen volume fraction inside the chamber may increase. Therefore, the control algorithm internally establishes a "negative pressure-flow rate" mapping curve, reads the output of the absolute differential pressure sensor in real time, and converts pressure fluctuations into additional flow rate compensation. Superimposed on the above This forms a dual-variable regulation, which can maintain low oxygen protection even when negative pressure fluctuates and material load changes drastically.

[0056] Example 3: In an aluminum powder coating workshop, the silo volume was 25 cubic meters, the negative pressure was set at 1800 Pa, and the oxygen volume fraction threshold was 13%. The carbon dioxide to nitrogen gas integral ratio was set to 1:3. During the trial operation, when the spray gun model was switched, the powder feed rate increased instantaneously, and the oxygen volume fraction rose from 12% to 12.8%. The feedforward term detected a 0.8% increase, and the control device increased the total flow rate of the mixed gas from 95 liters per minute to 140 liters per minute, bringing the oxygen volume fraction back to 12.2% in 20 seconds. The entire process was conducted without human intervention, and the negative pressure remained stable between 1750 Pa and 1850 Pa.

[0057] Example 4: In the test chamber of the magnesium powder recovery system, the top inspection port was intentionally opened for 2 seconds to simulate a leak. The negative pressure inside the chamber dropped by 200 Pa, and the oxygen probe detected a 0.5% increase in concentration. The negative pressure compensation function output an additional nitrogen flow rate of 30 liters per minute, and then restored the negative pressure and oxygen volume fraction to the set values ​​within 60 seconds. After the inspection port was closed again, the controller automatically reduced the flow rate to avoid excessive consumption of inert gas.

[0058] Continuous injection of inert gas mixture not only dilutes the oxygen volume fraction but also positively impacts the system's thermal balance. The high specific heat of carbon dioxide reduces heat dissipation when localized exothermic reactions occur in the filter bag or cyclone separator, while nitrogen reduces airflow viscosity, maintaining dust transport efficiency in the negative pressure airflow. Actual measurements show that continuous injection of inert gas under constant operating conditions can reduce the temperature of the outer wall of the metal casing by 4 to 6 degrees Celsius, reducing weld stress caused by thermal expansion and contraction.

[0059] To ensure reliable long-term operation, this invention incorporates an automatic drainage structure at the inlet of the gas mixing system to prevent residual moisture from the carbon dioxide cylinder from entering the mass flow meter and causing corrosion. The control software includes a flow meter zero-point self-check function, automatically closing the carbon dioxide channel and opening the nitrogen channel for pressure purging during shutdown, thus extending the lifespan of the probe and valve body.

[0060] like Figure 2 As shown, under the low oxygen conditions, the adjustable guide angle cyclone separator, the dielectric barrier discharge plasma reactor, and the waste heat driven thermoacoustic transducer are operated in coordination. The microphone array measures the sound field and the compressed sensing method is used to reconstruct the sound node amplitude. The sound node amplitude is combined with the outlet dust concentration and the combustible gas concentration to form the captured data.

[0061] Under low-oxygen conditions, this invention achieves simultaneous capture of combustible dust and associated combustible gases through a three-field coupling of gas, sound, and electricity. The core components consist sequentially of an adjustable-angle cyclone separator, a dielectric barrier discharge plasma reactor, and a waste heat-driven thermoacoustic transducer, which together form a synergistic capture unit. The cyclone separator provides a high-speed rotating flow field, the plasma reactor generates plasma charge and active free radicals, and the thermoacoustic transducer outputs acoustic energy and establishes a standing wave field at a fixed frequency. Its synergistic mechanism can be divided into four stages.

[0062] The first stage involves jet induction and particle size distribution adjustment. Dust-airflow enters the cyclone separator via guide vanes. The guide angle is adjusted in real-time by a servo motor to match the turbulent shear rate with the subsonic rotational speed at the acoustic field node position. The adjustment principle is based on feedback from the outlet dust concentration. If the concentration increases, indicating that large particles have not been completely stripped, the controller increases the guide angle to enhance centrifugal force. If the acoustic node amplitude decreases, the cone ratio is reduced to expand the acoustic cavity volume, maintaining the sound pressure standing wave resonance condition. This stage achieves coarse classification and acoustic field matching.

[0063] The second stage involves low-temperature plasma charging and gas-phase oxidation. The dielectric barrier discharge plasma reactor is installed on the outer wall of the cyclone separator near the outlet recirculation zone, with a ceramic dielectric between the electrodes. Electrons, negative oxygen ions, and active oxygen atoms are generated instantaneously during discharge. On one hand, electrons adsorbed on the particle surface acquire negative charges, and under the combined effects of rotating flow and acoustic pressure nodes, the charged particles attract each other and condense into larger particles. On the other hand, active oxygen atoms rapidly react with carbon monoxide or low-chain hydrocarbons in the combustible gas to generate carbon dioxide and water vapor, reducing the concentration of combustible components in the gas phase and providing fundamental data for subsequent free energy calculations.

[0064] The third stage is thermoacoustic-driven standing wave coalescence. Waste heat generated during the operation of the dielectric barrier discharge plasma reactor enters the thermoacoustic transducer through a heat conduction channel. After a temperature difference is established within the thermoacoustic transducer, a stable acoustic wave is excited. This acoustic wave is coupled into the cyclone separator cavity through an acoustic duct. The matching relationship between the cyclone length and the acoustic wavelength is utilized to form... The standing wave field generates high acoustic energy density at the sound pressure nodes, causing charged particles to rapidly converge under the combined action of steady-state acoustic radiation force and centrifugal force. The standing wave frequency is fixed, and the amplitude is adjusted according to the amount of waste heat, without consuming additional electrical energy, thus achieving integrated energy recovery and acoustic dust collection.

[0065] The fourth stage involves online sound field measurement and data synthesis. The microphone array is arranged at a non-uniform angle along the outer wall of the cyclone separator, reducing the number of mounting holes while maintaining sparse sampling. After each microphone outputs a sound pressure level signal, it is pre-amplified and converted from analog to digital before being sent to the edge computing unit. The edge computing unit records all sound pressure level sample vectors as... The measurement matrix is ​​denoted as The sound field modal coefficients are denoted as According to compressed sensing theory, sparse reconstruction satisfies:

[0066]

[0067] in For the observation vector, For the pre-calibrated sensing matrix, These are the sparse modal coefficients to be determined. The orthogonal matching pursuit algorithm can be used to find them within tens of iterations. Then calculate the amplitude of the acoustic node. .

[0068] After completing the acoustic node amplitude reconstruction, the system synchronously reads the outlet dust concentration output by the light scattering instrument. Combustible gas concentration output by the gas sensor The three data items are combined as follows:

[0069]

[0070] Pushed to the control flow within the control cycle. Here... Indicates the amplitude of the acoustic node. Indicates the dust concentration at the outlet. This indicates the concentration of combustible gas.

[0071] This invention improves the removal efficiency of submicron metal dust (diameter less than 2 micrometers) through multi-field coupling using a collaborative collection unit. Experimental results show that without acoustic field and plasma, the dust concentration at the cyclone outlet is approximately 180 mg / m³; after activating collaborative collection, the dust concentration at the outlet drops to 45 mg / m³, increasing the removal rate by approximately 75%. Simultaneously, the combustible gas concentration decreases by 60%, demonstrating the significant oxidation and dilution effects of plasma.

[0072] Example 5: In an aluminum powder shot blasting production line, the cyclone separator handles an air volume of 3000 cubic meters per hour, with an initial guide angle of 35 degrees and a cone ratio of 1.5. During operation, when the acoustic node amplitude reconstructed by the compressed sensing algorithm drops to 80% of the original set value, the controller adjusts the guide angle to 38 degrees and reduces the cone ratio to 1.4, restoring the acoustic node amplitude within 30 seconds; simultaneously, the outlet dust concentration drops from 60 mg / m³ to 48 mg / m³. This adjustment does not rely on manual interpretation, achieving adaptive dust collection.

[0073] Example 6: In a magnesium powder screening workshop, the thermoacoustic transducer output frequency was 22 kHz, and the acoustic node amplitude was 160 dB. The plasma reactor discharge power was 600 W. The system updates the captured data and transmits it to the control process with a sampling period of 0.5 seconds. During six months of continuous operation, no dust concentration exceeding the limit alarm occurred, proving that the capture mechanism established by the "acoustic-electrical-current" coupling in this invention has long-term stability.

[0074] Preferably, the microphone array is arranged at a non-uniform angle along the outer wall of the adjustable guide angle cyclone separator, and the sound pressure data acquired by the microphone array is sparsely reconstructed using the compressed sensing method to obtain the sound node amplitude.

[0075] To accurately obtain the spatial distribution of the sound field within a high-speed rotating, adjustable-angle cyclone separator, this invention employs a microphone array arranged at a non-uniform angle and combines it with a compressed sensing sparse reconstruction method to acquire the amplitude of sound nodes in real time. The cyclone cavity is approximately cylindrical, and sound waves are injected through a thermoacoustic transducer driven by waste heat, forming a fixed-frequency standing wave field within the cavity. Traditional uniform ring microphone layouts require a large number of measurement points to meet the Nyquist sampling criterion, which increases both the need for perforation and the weakening of the wall strength, as well as maintenance costs. This invention utilizes the prior characteristic that the sound modes within the cyclone cavity exhibit a sparse distribution in the angular direction, and introduces compressed sensing theory, allowing the reconstruction of the main sound field modes with only a small number of sensors.

[0076] The microphone array layout strategy is as follows: First, the angular distribution of sound pressure is derived based on the cyclone radius and standing wave frequency theory, and several priority sampling angles are pre-divided on the cavity wall. Then, a set of measurement points is generated using a random selection method of non-adjacent angles, ensuring that the measurement matrix satisfies the condition that the columns are approximately uncorrelated. The final array size accounts for approximately one-third of the uniform layout scheme. Each microphone is a condenser microphone, and its sensitivity is calibrated and stored locally in a lookup table. The sensor housing is welded to the outer wall through a threaded short tube, the length of which is controlled to be within one-twentieth of the incident wavelength of the sound wave to reduce the insertion effect.

[0077] After hardware data acquisition is complete, the software computing phase begins. Observation vectors are defined. Define a column vector of sound pressure level amplitudes output by the microphone array within a single sampling period, with dimensions equal to the number of microphones; define a measurement matrix. The sensor-modal response coefficient matrix is ​​obtained from acoustic experiments, with the number of columns equal to the number of possible acoustic modes; a sparse coefficient vector is defined. This represents the amplitude of each acoustic eigenmode at the current moment. The three satisfy the following relationship:

[0078]

[0079] in Represents the observation vector. Represents the measurement matrix. This represents the sparse modal coefficients. Because the actual sound field is dominated by a few low-order modal energies, It exhibits sparse characteristics. The algorithm employs orthogonal matching pursuit iterative selection. The column vector minimizes the residual norm, thus approximating the true modal coefficients. After sparse reconstruction, the amplitude of the mode corresponding to the standing wave node is selected and denoted as... That is, the amplitude of the acoustic node.

[0080] To suppress industrial noise interference, data preprocessing involves Hanning window smoothing in the time domain, followed by extraction of the target frequency band amplitude using Fast Fourier Transform, and then input into the compressed sensing module. The algorithm runs on an edge processor, with a single iteration taking less than 5 milliseconds, meeting the 0.2-second control cycle requirement. The average signal-to-noise ratio of sparse reconstruction is improved by approximately 12 dB, ensuring stable calculation of acoustic node amplitudes.

[0081] Nodal amplitude and outlet dust concentration Combustible gas concentration Together, they constitute the captured data. The controller, based on... Adjust the cyclone guide angle and cone ratio: When A decrease indicates energy attenuation of the standing wave field; increasing the guide angle increases centrifugal force while simultaneously reducing the cavity resonant length; conversely, when... If the set upper limit is exceeded, the guide angle should be appropriately reduced to prevent excessive acoustic energy accumulation and powder rebound. Within the same cycle, if... or Instead of decreasing, the controller will further increase the plasma discharge power and reduce the cyclone cone ratio to enhance charge coagulation and centrifugal separation effects.

[0082] Example 7: A 600 mm diameter cyclone separator was equipped with 12 microphones, 4 in the upstream recirculation zone and 8 in the downstream main separation zone. The standing wave frequency was set to 22 kHz. Test results showed that the frequency calculated after compressed sensing reconstruction... The error compared to the uniform 32-point sampling reference value is less than 7%. When the powder feed rate is increased to 120% of the rated value... The concentration rose to 65 milligrams per cubic meter, which was detected by the controller. The flow rate decreased by 15%, and within 20 seconds, the guide angle was automatically adjusted by 3 degrees and the plasma power was increased by 10%. It dropped back to 48 milligrams per cubic meter.

[0083] Example 8: Cyclone drum diameter 400 mm, equipped with 8 microphones. Customer requirement: Dust explosion index controlled to not exceed 0.3 kW / s per cubic meter. Data collected after 30 days of operation, after compressed sensing correction... The fluctuation range was maintained within ±8% of the set value, and the dust explosion index was below 0.25 kilowatt-seconds per cubic meter, achieving the safety production target.

[0084] The non-uniform angular arrangement combined with compressed sensing sparse reconstruction reduces the number of perforations, thus increasing the strength of the cylinder; it also reduces the number of sensors and wiring costs. Compared to the traditional large-scale uniform arrangement scheme, this invention reduces the number of sensors by approximately 60%, while increasing the measurement error of the acoustic node amplitude by no more than 10%, which can meet the requirements for fine control of dust collection. In addition, since numerical reconstruction is completed only in the edge processor, there is no need to upload the entire acoustic pressure waveform to the central server, avoiding high-bandwidth data flow occupation and improving system real-time performance.

[0085] Preferably, the adjustable guide angle cyclone separator synchronously adjusts the guide angle and cone ratio through an electric actuator. The guide angle and cone ratio are adjusted in a closed loop according to the real-time changes in the acoustic node amplitude and the outlet dust concentration.

[0086] In negative pressure circulating explosion-controlled dust removal systems, cyclone separators not only perform conventional centrifugal classification but also coordinate with standing wave sound fields. To adapt to dynamic fluctuations in dust concentration, acoustic energy, and airflow, this invention designs the cyclone separator with a dual adjustable structure: the guide vane angle is used to adjust the inlet tangential momentum, and the cone ratio is used to adjust the particle residence path on the wall. These two geometric parameters are labeled as the guide vane angle. Compared to a cone Both are driven by independent electric actuators, but are adjusted synchronously under the same control law, so that the centrifugal force field and the acoustic standing wave field always maintain optimal matching.

[0087] In terms of structure, the guide vane assembly is installed inside the spiral duct at the cyclone inlet, and the guide vanes are directly connected to the servo motor via a stainless steel central shaft. The angular displacement of the motor rotor is obtained in real time via a magneto-electric encoder. The conical section is connected to the cylindrical section via a linear slide rail. The ball screw on the outer wall is driven by a stepper motor, enabling smooth axial extension and retraction of the conical section. A slide rail displacement sensor outputs the real-time extension and retraction of the conical section, which is then geometrically calculated to obtain... .

[0088] The controller collects the amplitude of the acoustic nodes every 0.2 seconds. (Reconstructed from compressed sensing by microphone array) and outlet dust concentration (Measured by a laser scattering instrument). Both and the preset target. , After comparison, the system proceeds to dual-channel proportional-integral (PI) calculation. The control law is represented in matrix form:

[0089]

[0090] in Indicates the correction amount (degrees) for the guide angle; Indicates the cone ratio correction amount; , Indicates the proportionality coefficient; , These represent the integral coefficients. The four coefficients were tuned experimentally using a step response to ensure the system converges within five control cycles without overshoot. , Applying exponential smoothing yields the command angle increment. With expansion and contraction displacement Then, the data is written to the two motor drivers via the EtherCAT bus. (Smoothness factor) It can be adjusted via the touchscreen interface.

[0091] In principle, when If the value is lower than the target value, it indicates that the acoustic standing wave energy is insufficient, and the system will prioritize increasing it. To increase the tangential velocity, the acoustic cavity length is slightly shortened and shifted towards higher-order modes, thereby increasing the standing wave sound pressure; if adjusted... If the safety voltage drop is still met, then make a fine adjustment. Improve centrifugal collection efficiency. When Higher than the target value Normal, the controller mainly reduces (Elongation of the cone section) extends the particle sliding path along the wall, increasing the probability of particle dust settling, while slightly increasing... This compensates for resonance deviations caused by changes in the acoustic cavity length. The linkage of the two parameters avoids the increased energy consumption that would result from relying solely on airflow or acoustic power for compensation.

[0092] The technical effectiveness has been verified through long-term experiments. After 60 days of continuous operation on a 0.6-meter diameter cyclone separator, compared to a fixed geometry structure, the submicron dust removal efficiency increased from 78% to 89%. The amplitude fluctuation of the acoustic node decreased from ±15% to ±6%. Pressure loss increased by only 7%, lower than the over 20% increase of conventional throttling compensation schemes. Automatic adjustments occurred only 3% of the total operating time, indicating that the control system is stabilizing.

[0093] Example 9: Aluminum powder production line with a gas processing capacity of 3200 cubic meters per hour. Initial setup. Spend, When the dust load suddenly increases by 40%, It jumped to 70 milligrams per cubic meter. A decrease of 18%. The controller will first... Adjust the elevation of the sound pressure point to 33 degrees, then further reduce the cone ratio to 1.45. .final Revert to 2% above the target value. The pressure dropped to 48 mg / m³. The pressure fluctuation throughout the process was less than 90 Pa, and the emergency stop threshold was not triggered.

[0094] Preferably, the dielectric barrier discharge plasma reactor adjusts the voltage between electrodes and the electrode spacing to match the discharge intensity with the oxygen volume fraction, thereby maintaining a stable ionization state and suppressing spark generation.

[0095] The dielectric barrier discharge plasma reactor is located on the outer wall of the reflux zone of a cyclone separator, employing a coaxial cylindrical structure: the outer electrode is a thin-shell stainless steel shell, the inner electrode is a copper alloy rod, and a uniformly thick alumina dielectric tube is sandwiched in between. When operating in a low-oxygen environment, this reactor needs to simultaneously meet two requirements: first, maintaining uniform plasma volume discharge without single-point arcing; and second, providing sufficient plasma density to inject charge into the surface of fine dust particles and generate active oxygen atoms, further oxidizing combustible gases. Since this invention dilutes the oxygen volume fraction within the chamber using an inert gas mixture, the oxygen partial pressure changes in real time with the operating conditions. Maintaining a fixed electrode spacing and a fixed peak voltage can easily lead to excessive or insufficient breakdown field strength. Therefore, this invention proposes a "voltage-electrode spacing dual-channel adaptive adjustment" scheme, dynamically matching the discharge intensity with the oxygen volume fraction, ensuring sufficient ionization while eliminating the risk of sparks.

[0096] The discharge matching principle states that gas breakdown voltage is closely related to gas pressure, electrode spacing, and gas composition. For a low-oxygen mixture of air, carbon dioxide, and nitrogen, the breakdown voltage can be approximated using a modified Pascal-Schönlein relationship. With oxygen partial pressure and electrode spacing Functions:

[0097]

[0098] Indicates the breakdown voltage; This represents the thickness of the dielectric layer plus the air gap thickness; Indicates the current oxygen partial pressure; This constant, related to the mean free path of gas molecules and the electron collision cross section, is fixed after calibration of this device as the oxygen volume fraction decreases. If the original electrode spacing and voltage are maintained, the electric field coefficient will decrease. Above the breakdown threshold, an excessively strong local electric field can create a thermal arc, igniting the dust. Conversely, if the oxygen volume fraction increases, an insufficient electric field will lead to a decrease in plasma density and a reduction in charging efficiency. Therefore, it is necessary to adjust the parameters in real time based on the oxygen partial pressure. or The actual electric field strength Discharge intensity close to the target .

[0099] The reactor's structure and actuator include an integrated ball screw-servo motor moving actuator on the outer shell, driving the outer electrode to make axial micro-movements relative to the inner electrode, thereby continuously adjusting the electrode spacing. A high-frequency, high-voltage power supply can rapidly increase and decrease within its peak voltage range via a digital modulator. The oxygen sensor outputs the oxygen volume fraction in 0.5-second cycles. The controller converts it into oxygen partial pressure. Then calculate the target voltage based on the target discharge intensity. Distance from target If the difference between the two values ​​and the current setting exceeds the set threshold, dual-channel adjustment is activated: first, the external electrode is slowly moved to... Nearby, fine-tune the high-voltage power supply output to ensure that it always meets the requirements during the process. .

[0100] Control algorithm, assuming the current sampling number is First, calculate the error vector:

[0101]

[0102] in Calculated using the Paschen formula, , These represent the current peak voltage and the electrode spacing, respectively. Then, applying the proportional-integral matrix control law:

[0103]

[0104] and for A diagonal matrix, with coefficients experimentally tuned. Control increment. After amplitude limiting, slope limiting, and dead zone processing, the data is written to the power supply buck-boost module and the servo movement module respectively. The buck-boost slope is limited to within 100 volts per millisecond; the electrode movement speed does not exceed 2 millimeters per second. This strategy ensures that the reactor maintains a uniform glow discharge state even when oxygen partial pressure fluctuates or inert gas suddenly increases.

[0105] Through this invention, uniform glow discharge generates a plasma density of approximately 10. 11 Each cubic centimeter is sufficient to allow a large number of electrons to be adsorbed onto the surface of submicron dust within tens of milliseconds, forming a negative charge; simultaneously, bombarding air and carbon dioxide molecules produces reactive oxygen atoms and ozone, which can react with carbon monoxide and low-chain hydrocarbons in microseconds. Continuous spectral monitoring shows that the discharge spectrum under the adjustment strategy of this invention is dominated by the arc band of secondary ionized nitrogen molecules, and no high-intensity hydrogen radiation lines were detected, indicating the absence of a high-temperature arc. Power measurements prove that within the oxygen volume fraction range of 12% to 18%, the total discharge power fluctuation is less than ±7%, and the discharge efficiency remains stable.

[0106] Example 10: An aluminum powder shot blasting production line with a dust collection chamber volume of 30 cubic meters and an initial oxygen volume fraction of 17%. The target discharge intensity was set at 4.8 kV / cm, the initial electrode spacing at 2.5 mm, and the peak voltage at 12 kV. After the system activated the inert gas mixture, the oxygen volume fraction dropped to 13% within 60 seconds. The controller detected a decrease in the breakdown voltage threshold, first reducing the peak voltage to 10 kV, then increasing the electrode spacing to 3.2 mm. No sparks were heard throughout the process, and the monitoring current remained stable at 0.38 mA. After two weeks of continuous operation, the average dust charging efficiency increased by 12%, and the outlet combustible gas concentration decreased from 110 mg / m³ to 45 mg / m³.

[0107] Preferably, the waste heat-driven thermoacoustic transducer receives waste heat generated by the dielectric barrier discharge plasma reactor and releases acoustic energy at a fixed acoustic frequency to enhance the coagulation of fine dust particles.

[0108] The waste heat-driven thermoacoustic transducer utilizes the natural temperature difference between the outer shell and cooling sleeve of the dielectric barrier discharge plasma reactor to convert passively dissipated heat into fixed-frequency acoustic energy, thereby enhancing the acoustic coagulation process of fine dust particles. The transducer consists of four parts: a hot-end heat exchange cavity, a stacked structure, a cold-end heat exchange cavity, and a resonant tube. The hot-end heat exchange cavity is fitted to the outer wall of the plasma reactor and secured to the reactor shell by a high-thermal-conductivity copper plate; the cold-end heat exchange cavity is maintained at a low temperature by a water-cooled jacket. The stacked structure uses ceramic honeycomb blocks with pore sizes matched to the Prandtl number, enabling effective heat transfer during the cyclic compression and expansion of thermoacoustic waves between the solid wall and the air column, thus driving the self-excitation of the acoustic waves. The length of the resonant tube is calculated so that the lowest-order longitudinal mode falls around 22 kHz, the same frequency as the standing wave frequency of a cyclone, which can be coupled into the cylinder through an acoustic duct.

[0109] The thermoacoustic start-up criterion follows one-dimensional thermoacoustic theory, when the hot end temperature... With cold end temperature The difference exceeds the critical temperature difference The sound wave will spontaneously increase. The critical temperature difference can be written as:

[0110]

[0111] in The thickness coefficient of the Stokes layer. The average temperature. This refers to the specific heat ratio. The plasma reactor shell temperature is typically maintained around 180 degrees Celsius, while the cold-end jacket water temperature is kept at 35 degrees Celsius. The actual temperature difference is greater than [missing value]. Approximately 20% ensures the transducer remains in its self-excited region across the entire load range. Acoustic power output. With heat flow The following conditions must be met:

[0112]

[0113] The acoustic-to-thermal conversion efficiency was experimentally calibrated to be approximately 0.18. The heat source was solely waste heat from the plasma reactor, with no additional electrical energy consumption. Consequently, the sound pressure level remained stable at around 160 dB, with a fluctuation bandwidth of ±3 dB, meeting the energy threshold for acoustic coagulation of submicron dust.

[0114] Sound waves are coupled into the inner cavity of the cyclone separator through a 25 mm diameter acoustic duct. Since the length ratio of the resonant tube to the cyclone separator is approximately an integer multiple, a sound wave is formed inside the separator after coupling. Standing waves, with the node position fixed at the axis of rotation. Charged fine dust particles accumulate at the node driven by both acoustic radiation force and electrostatic Coulomb force, and are subsequently flung towards the wall by enhanced centrifugal force. Acoustic radiation force The sound pressure amplitude is directly proportional to the square of the sound pressure, and the sound pressure amplitude is directly controlled by the waste heat power. When the reactor load fluctuates and the shell temperature drops, the sound amplitude will drop accordingly. The controller compensates for the centrifugal force by reducing the cone ratio and increasing the guide angle to maintain the overall capture efficiency.

[0115] To avoid frequency drift in the thermoacoustic transducer output, this invention incorporates a miniature piezoelectric transducer within the resonant tube as a phase-locked probe. The controller continuously monitors the phase difference between the transducer's potential phase and the acoustic pressure phase at the acoustic duct inlet. If this difference exceeds 15 degrees, the length of the threaded microtube plug in the middle of the resonant tube is finely adjusted by ±0.5 mm to bring the resonant frequency back to the target value. This structure achieves precise frequency locking without altering the hot-end temperature difference, ensuring consistency with the measurement matrix parameters used in microphone arrays.

[0116] In Example 11, in a magnesium powder screening workshop, the plasma shell heat flux was 2.5 kW, the measured acoustic power of the thermoacoustic transducer was 450 W, and the sound pressure level was 158 dB. After continuous operation for 8 hours, the outlet dust concentration decreased from 65 mg / m³ to 42 mg / m³; after the transducer was turned off, the concentration rose back to 60 mg / m³, confirming that the acoustic field contributes approximately 28% to the dust removal effect on fine particle agglomeration.

[0117] In Example 12, during the aluminum powder recovery system test, the waste heat fluctuation cycle was 300 seconds, with a peak-to-valley difference of 25 degrees Celsius. The transducer output sound pressure fluctuation was ±2.5 dB, but through cyclone geometric synchronization adjustment, the outlet dust concentration fluctuation was limited to ±7%. High-temperature infrared imaging showed that the transducer hot end temperature reached a maximum of 190 degrees Celsius, and the cold end temperature reached a maximum of 38 degrees Celsius. The aluminum alloy shell showed no fatigue cracks, indicating that the temperature difference cycle was within the material's allowable range.

[0118] The environmental data and the captured data are combined to form a state vector, which is then input into a control flow containing a differential thermochemical model and a reinforcement learning network to generate and issue control commands for adjusting negative pressure pulses, inert gas flow rates, sound field intensity, and cyclone geometry parameters.

[0119] Environmental data comes from pressure, oxygen content, and temperature-strain sensor networks, while data acquisition comes from acoustic node amplitude, outlet dust concentration, and combustible gas concentration acquisition links. The controller first concatenates the two types of data in a fixed order to obtain a nine-dimensional state vector:

[0120]

[0121] The warehouse is under negative pressure. This refers to the oxygen volume fraction. The average temperature of the wall surface. For the maximum strain of the wall, The amplitude of the sound node. The concentration of dust at the outlet. Combustible gas concentration, and These represent the guide angle and cone ratio, respectively. All quantities are subjected to min-max linear normalization to distribute data of different dimensions within the interval [0, 1], avoiding numerical explosion while preserving relative amplitude information. The control period is set to 0.2 seconds, and the vector maintains the parameter meaning and order within the period to facilitate subsequent gradient propagation.

[0122] To enable intelligent algorithms to not only "see" measured values ​​but also perceive potential explosion risks, this invention inserts a differential thermochemical model at the front end of the policy network. The model expresses free energy and exothermic rates as differential thermochemical parameters. Find the analytical form of the gradient. Assume the molar amount of combustible gas in the system. Molar amount of oxygen absolute temperature The instantaneous rate of the main exothermic reaction is expressed according to the Arrhenius equation:

[0123]

[0124] in For frequency factors, For activation energy, Let be the gas constant. System free energy:

[0125]

[0126] For the first The chemical potential of the components Its molar quantity. Because , as well as All can be In , and A unique mapping can be obtained using automatic differentiation. and The gradient serves as a physical regularization term in subsequent reinforcement learning training, ensuring that the policy update direction is consistent with "reducing free energy and suppressing heat release," and preventing it from crossing the safety boundary due to black-box exploration.

[0127] The policy network employs a three-layer architecture: graph convolution, long short-term memory (LSTM), and fully connected layers. The first layer constructs six physical actuators—the fan, mixing valve, guide vane mechanism, cone mechanism, thermoacoustic transducer, and plasma power supply—as graph nodes, with edge weights defined by energy coupling coefficients. Graph convolution extracts spatial coupling features. The second layer uses two layers of LSM to process the state sequence of the most recent ten frames, capturing dynamic trends. The third fully connected layer outputs action vectors.

[0128]

[0129] in The amplitude of the negative pressure pulse to be applied. The target flow rate for the inert gas mixture. The target amplitude of the sound field. , The target geometry is a cyclone. Network training uses a proximal policy optimization, with the following overall loss:

[0130]

[0131] , The weights are determined via a grid search. and It can be transmitted back, and the network converges after 3000 iterations, which is 35% faster than the control group without physical regularization, and the strategy touches the free energy peak region less often. The action vector, after being processed by a safety limiter, is sent out in parallel via the EtherCAT bus: the inverter receives... Adjusting the fan torque generates short negative pressure pulses within the 100 to 250 millisecond range; the mass flow controller... Regulate the carbon dioxide-nitrogen mixture flow; programmable power amplifier drives the thermoacoustic transducer output. The servo motor and stepper motor respectively position the guide vane and the cone to... , The actual position, flow rate, voltage, and other feedback quantities of all actuators are packaged into environmental data or captured data for the next cycle, realizing a closed loop of "measurement-calculation-control-verification".

[0132] Statistical results from 60 consecutive days of industrial operation show that after adopting this control process, the average submicron dust emission was 41 mg / m³, a 30% reduction compared to traditional control methods; the average concentration of combustible gases decreased by 58%; and the system's energy consumption per unit air volume increased by only 6%. In three simulated leak tests, after the oxygen volume fraction rapidly rose to 15%, this process output a combination of maximum inerting flow and high-amplitude negative pressure pulses within 0.4 seconds, pulling the oxygen volume fraction back below 12% within 2 seconds, while maintaining resonance in the acoustic field within the cyclone and preventing secondary dust re-suspension. The measured peak free energy decreased by 22%, verifying the explosion suppression effect of thermochemical canonical methods on extreme disturbances.

[0133] Example 13: With a processing capacity of 3200 cubic meters per hour, when the feed rate suddenly increases by 40%, the outlet dust concentration jumps to 70 milligrams per cubic meter, and the acoustic node amplitude decreases by 18%. After the state vector is input, the network outputs the following actions: Pa, Rise per minute decibel, , The fan, valves, and geometric mechanisms complete adjustment within 15 seconds. It fell back to 48 milligrams per cubic meter. The free energy decreased by 17% compared to before the impact, and the free energy after the impact did not exceed the safety threshold.

[0134] Preferably, a microthermochemical model can be used to calculate the explosion free energy and system entropy and output a gradient. A reinforcement learning network updates the control strategy based on the gradient to generate control commands for adjusting the negative pressure pulse, inert gas flow rate, sound field intensity, and cyclone geometry parameters.

[0135] In combustible dust environments, if free energy accumulates rapidly while entropy decreases slowly, spontaneous combustion or secondary detonation can be easily triggered. This invention introduces a differentiable thermochemical model at the control level, using "explosion free energy - system entropy" as the core safety indicator. By automatically differentiating, its gradient with respect to each state variable is directly transmitted to the reinforcement learning network, ensuring that the policy optimization direction is strictly consistent with thermodynamic safety criteria, rather than relying on empirical thresholds or single concentration rules.

[0136] The microthermochemical model first physically expresses the chemical potential energy of the system. Assume the main combustible component in the gas mixture is carbon monoxide, with a molar mass denoted as . The molar amount of oxygen is denoted as The absolute temperature of the system is denoted as Free energy, written according to the classical definition:

[0137]

[0138] in and These represent the chemical potentials of the corresponding components. The chemical potential can be obtained by converting the Gibbs-Duhem relation to the instantaneous partial pressure. With state vector The oxygen volume fraction, combustible gas concentration, and temperature in the system exhibit differentiability. The system entropy can be written as:

[0139]

[0140] For isobaric specific heat capacity, This is the gas constant. Because... and right All are explicitly differentiable functions, and their gradients can be obtained through automatic differentiation. and These reflect the sensitivity of free energy to various physical quantities and its contribution to entropy change, serving as first-hand "physical signals" for reinforcement learning strategies to assess environmental risk. Reinforcement learning networks employ a proximal policy optimization framework. At each time step, the network observes the current state vector. And output the action vector —That is, the negative pressure pulse amplitude, inert gas flow rate setting, sound field intensity setting, and the target cyclone guide angle and cone ratio. Strategy parameters The update objective is not only to maximize the time discount reward, but also to minimize the free energy and maximize the entropy; therefore, a composite loss function is adopted:

[0141]

[0142] , The weights are used for the gradient update. The gradient update then includes the usual policy gradient term and the physical gradient term from the differentiable thermochemical model:

[0143]

[0144] In the above formula, " " is the vector inner product, and Provided directly by a differentiable model, This is then obtained through the backpropagation chain of the network. In this way, if a certain state component causes a sharp increase in free energy or a sharp decrease in entropy, its corresponding gradient will increase, exerting a stronger constraint on the policy parameters. This makes the network more inclined to reduce the dangerous component in the next round of output actions, such as increasing inertial gas injection or amplifying the negative pressure pulse.

[0145] To ensure the actions remain within the executable range, this invention incorporates amplitude and slope constraints at the output: the negative pressure pulse amplitude does not exceed 0.3 times the rated differential pressure of the fan, the instantaneous inert gas injection rate does not exceed 100 liters per minute, the target sound field amplitude does not exceed the safe power of the thermoacoustic transducer, and the cyclone geometry parameter adjustment rate is limited to within 3 degrees per second and a scaling ratio of 0.05. The controller writes the finalized instructions in parallel to the drivers of each subsystem via the EtherCAT bus, achieving millisecond-level execution.

[0146] Example 14: Simulated 60% surge in feed rate on aluminum powder coating line: Free energy increases by 25% within 0.4 seconds. Reinforcement learning network perceives this. Upon a sudden increase in oxygen volume fraction and dust concentration, a maximum negative pressure pulse of 300 Pa and an inert gas flow rate increase of 80 liters per minute are immediately output, along with a 4 dB increase in sound field amplitude, a 35-degree guide angle, and a decrease in cone ratio to 1.45. After 10 seconds, the free energy decreases by 30% from its peak value, and both dust and combustible gas concentrations return to safe levels. If the physical gradient constraint is removed, relying solely on… During training, the policy lag was approximately 2 seconds in the same scenario, and the maximum free energy exceeded the safety threshold.

[0147] Preferably, the reinforcement learning network includes a graph neural network for describing the coupling relationships between devices within the dust removal system, and a time-series encoder for processing the time-series features of environmental data and captured data.

[0148] During the operation of the dust removal system, multiple couplings exist between negative pressure pulses, inert gas injection, cyclone geometry changes, and acoustic energy: changes in the fan's suction volume affect the pressure and oxygen volume fraction within the chamber; changes in oxygen volume fraction, in turn, determine the dielectric barrier discharge power and thermoacoustic waste heat; and acoustic energy attenuation must be compensated for by adjusting the guide angle and cone ratio. To comprehensively consider these interactions within a 0.2-second control cycle, the traditional method of simply concatenating all measurements into a vector and then inputting it into a multilayer perceptron is inefficient and has poor adaptability to new materials or different production lines. This invention introduces a dual-stream structure of "graph neural network-temporal encoder" in the encoding stage of the reinforcement learning policy network: the graph neural network explicitly describes the equipment topology and energy coupling, while the temporal encoder extracts environmental data and the dynamic evolution of captured data, thereby simultaneously leveraging the advantages of the attention mechanism in both spatial and temporal dimensions.

[0149] Spatial modeling and graph neural network implementation: The system selects six core execution units as nodes, namely the fan, mass flow valve, guide vane servo mechanism, cone stepper mechanism, thermoacoustic power amplifier, and plasma power supply, denoted as set. Directed edges are established based on the direction of energy or matter transfer, for example, an edge is established from the fan to the mass flow valve. Mass flow valve to cyclone separator edge The energy coupling coefficient is measured as the percentage of power transferred per unit time or volumetric flow rate and is written into the adjacency matrix. Initial feature vector of a node It consists of three parts: the real-time operating conditions of this node (such as motor speed and power supply voltage), and a state vector segment (such as...). ) and normalized device health metrics. The convolutional layer uses normalized message passing:

[0150]

[0151] in Normalize, It is a linear rectified function. This is a bias. After three layers of propagation, the node embeddings with fused neighborhood information are obtained. Then, average pooling is used to obtain the global space vector. .

[0152] Since the adjacency matrix elements are derived from actual energy coupling, if the type of dust is changed or the process is altered, only the edge weights need to be fine-tuned, without the need to redesign the network structure, thus ensuring the portability of the strategy.

[0153] Temporal modeling and a temporal encoder are implemented, as parameters such as fine dust concentration, acoustic node amplitude, and oxygen volume fraction exhibit strong time delays. To capture cross-period dependencies, the encoder employs a two-layer long short-term memory network, inputting a sequence of state vectors from the most recent 10 frames. The hidden dimension is 64. The memory gate controls information that decays within a 2-second window, while the forget gate adaptively adjusts based on the free energy gradient, allowing potential explosion trends to be exposed earlier. The final output is the hidden state. The system dynamics of the previous stage are summarized.

[0154] Fusion and motion generation, and Concatenate the following: Input a three-layer fully connected network, output action vectors.

[0155]

[0156] in Unit Pa, Unit: liters per minute Unit: decibel Unit of degree The ratio is a dimensionless cone. Each component first passes through a limiter, then is sent to the corresponding actuator via an EtherCAT bus. Network training employs a proximal-end policy optimization, superimposed with physical regularization terms for free energy and entropy. The loss function is:

[0157]

[0158] The strategy is determined through grid search, which allows the strategy to both pursue high returns and comply with thermal safety constraints.

[0159] In Example 15, on a production line with a flow rate of 3500 cubic meters per hour and a cyclone diameter of 600 millimeters, after adding graph-time coding, the number of convergence samples for the strategy decreased from 15,000 to 10,000. When migrating to the titanium powder scenario, adjusting only two edges in A is sufficient to maintain the control level of 45 milligrams per cubic meter of dust at the outlet.

[0160] On-site statistics over 30 days showed that the average emission of submicron dust was 41 milligrams per cubic meter, and the concentration of combustible gas was reduced by 58% compared to the control measures. Under extreme load impact, the peak free energy of the explosion was reduced by 24%, and the system energy consumption increased by only 6%.

[0161] Example 16: Fluctuations in the cold-end water pressure of a thermoacoustic power amplifier caused a 15% instantaneous drop in sound power. The graph neural network first captured this. When node characteristics change, the changes propagate to the geometric nodes via edge weights. The strategy immediately issues a "increase the guide angle by 4 degrees and reduce the cone ratio by 0.08" command, which restores the acoustic node amplitude and maintains a low dust concentration within 15 seconds without manual intervention or shutdown.

[0162] like Figure 3 As shown, based on the environmental data, the control commands, and the radio frequency imaging data acquired by the distributed antenna array, the explosion free energy is predicted using a generative adversarial digital twin model. When the current or predicted explosion free energy exceeds the threshold, a strategy rollback command or an emergency shutdown command is sent to achieve closed-loop control.

[0163] The distributed antenna array consists of 32 microstrip transmit-receive elements uniformly arranged around the cyclone tube, operating in the frequency band from 4 GHz to 6 GHz. Each periodic antenna transmits short pulses in a time-division sequence, and the array controller synchronously acquires the echo amplitude and phase. A multipath-sparse reconstruction algorithm is used to obtain a voxel map of the instantaneous complex permittivity distribution inside the tube, denoted as the RF imaging data volume. The change in dielectric constant is highly correlated with local dust concentration, temperature gradient, and water vapor content. This enables 3D scene observation, which is impossible with traditional single-point sensors. The controller concatenates the radio frequency imaging object, environmental data vector, and the control commands issued in the previous step into an extended state tensor. It updates every 0.2 seconds. This tensor input generates an adversarial digital twin model: the generator employs a physically guided convolutional network, with the first 5 frames... The generator conditionally predicts the dust concentration field, oxygen volume fraction field, and temperature field within the next 1-2 seconds. A discriminator compares the generated field with the actual field measured by a network of radio frequency, oxygen content, and temperature-strain sensors, outputting an adversarial loss and updating the generator weights accordingly. Finally, the generator calculates the predicted explosion free energy in the tensor domain. Free energy can be expressed using the Gibbs relation:

[0164]

[0165] in For enthalpy, Absolute temperature Entropy is the specific enthalpy of the dust-gas mixture, obtained by integrating the predicted temperature field; entropy is calculated from the dust-gas fraction and local turbulent kinetic energy. The model can be used for... Automatic differentiation minimizes both adversarial loss and [other losses] during training. The time variance ensures a sensitive response to changes in free energy.

[0166] System preset security threshold When the measured free energy or predicted value Any one of them exceeds The risk control dispatcher immediately enters a two-level protection phase: strategy rollback, canceling the output of the two most recent reinforcement learning strategies, and resetting the negative pressure pulse, inert gas flow, sound field intensity, and cyclone geometry to the most recently verified and safe backup parameters; emergency shutdown (if within 0.5 seconds). If the value is still above the threshold, the system will trigger a power outage for the fan, maximum injection of inert gas, disconnection of the plasma power supply, instantaneous opening of the cyclone ash-blocking valve to relieve pressure, and issue a shutdown signal to the higher-level DCS.

[0167] This invention provides continuous spatial resolution for radio frequency volumetric imaging, enabling the capture of local enrichment and hotspot coupling in dust clouds, thus compensating for blind spots in single-point measurements. A generative adversarial digital twin model employs an online self-calibration method using a test-predictive adversarial approach, avoiding the inaccuracies caused by parameter drift in traditional physical simulation models. By directly incorporating safety indicators into the network gradient using the differentiable free energy formula, both model interpretability and training efficiency are maintained. A two-level protection strategy—rollback followed by shutdown—minimizes the probability of false shutdowns and production losses.

[0168] Example 17: Magnesium powder coating line, cylinder diameter 0.6m, RF voxel resolution 20mm. During a certain incident, the coating gun became clogged, and dust instantly accumulated in the upstream return zone. The peak dielectric constant is increased by 35%. The generator predicts the free energy after 0.8 s. The system executed a strategy rollback within 90ms: reducing the guide angle by 5°, increasing the inertial flow rate by 70L / min, and increasing the sound pressure by 4dB. The real-time measured free energy decreased to 0.83 within 2s. No shutdown was triggered. If the RF-twin module is disabled, relying solely on threshold control, the actual peak free energy reaches 1.3. Furthermore, it triggered an emergency shutdown, causing the production line to stop for 40 minutes.

[0169] Preferably, the distributed antenna array is deployed in the form of a planar array. It obtains radio frequency reflection amplitude and phase information by scanning the working area through phase modulation to form radio frequency imaging data. It generates an adversarial digital twin model, combines fluid dynamics calculation and discrete element calculation, and perturbs the Green tensor to simulate the most unfavorable working condition. When the current or predicted explosion free energy exceeds the threshold, the emergency shutdown command raises the negative pressure pulse and inert gas flow rate to the preset maximum safe value, and keeps the sound field intensity and the guide angle and cone ratio of the adjustable guide angle cyclone separator at the most recent safe state.

[0170] A distributed antenna array is arranged in a rectangular planar array along the outer wall of the dust removal chamber, with the element spacing set to half a wavelength, operating in the 4-6 GHz frequency band. The array controller assigns pre-calculated phase codes to adjacent antennas, which then transmit short pulses in turn according to a quasi-orthogonal sequence; the remaining antennas synchronously receive the echoes, and the amplitude-phase matrix is ​​obtained through cross-correlation. A three-dimensional complex permittivity voxel map is then obtained through sparse Fourier inversion and conjugate matched deconvolution. This radio frequency imaging data is particularly sensitive to local dust concentration, gas-solid temperature difference, and minute spark discharges, and can fill point sensing blind spots in real time.

[0171] The generative adversarial digital twin model outputs the dust concentration field, oxygen volume fraction field, and temperature field for the next second, based on the most recent five frames of RF voxel maps, synchronized environmental data, and control commands. The model consists of a "physics-guided generator + adversarial discriminator": the generator backbone is a 3D convolutional-cross-scale residual network, embedding a perturbation module based on Green's tensor to superimpose the most unfavorable perturbation onto the input field to approximate possible extreme distributions; the discriminator uses a PatchGAN structure to provide adversarial loss for the difference between the generated field and the measured field. During the training phase, a fluid dynamics-discrete element co-simulation loss is added. The generator must simultaneously minimize the adversarial error and the gap with the computational fluid dynamics-discrete element coupled simulation results to ensure that the predictions are both realistic and conservation-related.

[0172] The model calculates the explosion free energy using:

[0173]

[0174] In the formula For the first The chemical potential of the components These are molar quantities, obtained by integrating the predicted field. The system continuously compares the measured free energy. With predicted free energy If any value exceeds the threshold Immediately initiate a safety closed loop:

[0175] Emergency shutdown command: The frequency converter will increase the negative pressure pulse amplitude to the preset maximum safety value (the pressure difference is about 1.3 times the normal cruise value), and the mass flow controller will increase the carbon dioxide-nitrogen mixed flow rate to the rated limit to quickly dilute the oxygen volume fraction and remove sensible heat.

[0176] Freezing sound-cyclone parameters: The output of the thermoacoustic transducer amplifier is maintained at the sound pressure level that was recently verified to be safe, and the guide vane angle and cone ratio are locked at the last effective geometry to prevent adjustment disturbances from causing secondary dust re-suspension.

[0177] Instruction rollback record: The controller writes the policy parameters of the two most recent frames that caused the threshold to be exceeded into the black box for use in offline retraining and parameter recalibration.

[0178] This invention enables planar array phase scanning to cover the entire chamber volume without mechanical rotation, with a sampling period of 50ms, meeting the requirements for transient dust cloud detection. The Green's tensor perturbation incorporates worst-case scenario assumptions at the generation stage, ensuring the predicted envelope reflects real risk; adversarial training allows the model to self-correct even when the dielectric properties of new powder drift. The shutdown logic first rapidly increases the negative pressure and inert gas flow rate, then maintains the acoustic-cyclone combination in a safe state, quickly extinguishing the explosion while avoiding excessive perturbation that could lead to dust accumulation and re-suspension.

[0179] Example 18: In a magnesium powder production line processing 3200 m³ / h, a simulated momentary dust release caused by gate jamming in the feed hopper was conducted. The RF array detected a 28% increase in local dielectric constant within 70 ms, and the generator predicted that the free energy would reach 1.15 after 0.6 s. The system immediately triggered an emergency shutdown: the negative pressure pulse increased to 280 Pa, the inert flow rate increased by 90 L / min; within 2 seconds, the oxygen volume fraction decreased from 13% to 11%, the peak internal temperature dropped by 12°C, and the free energy decreased to 0.8. Production was uninterrupted. If the RF-twin module is shut down and only the point sensor is used, the actual peak free energy reaches 1.32 due to data delay. This triggered a 35-minute shutdown of the entire production line.

[0180] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A negative pressure circulating explosion-controlled dust removal method suitable for combustible dust environments, characterized in that, Includes the following steps: The sealed dust removal system creates negative pressure by suction and injects an inert mixture of carbon dioxide and nitrogen to lower the oxygen volume fraction below a safe threshold. It collects negative pressure value, oxygen volume fraction, and temperature-strain information to generate environmental data. The temperature-strain information is obtained by fiber Bragg gratings arranged in segments along the wall of the dust removal system. Under low-oxygen conditions, the adjustable guide angle cyclone separator, dielectric barrier discharge plasma reactor and waste heat driven thermoacoustic transducer are operated in synergy. The microphone array measures the sound field and the compressed sensing method is used to reconstruct the sound node amplitude. The sound node amplitude is combined with the outlet dust concentration and combustible gas concentration to form the captured data. The environmental data and the captured data are combined to form a state vector, which is then input into a control flow containing a differential thermochemical model and a reinforcement learning network to generate and issue control commands for adjusting negative pressure pulses, inert gas flow rates, sound field intensity, and cyclone geometry parameters. Based on the environmental data, the control commands, and the radio frequency imaging data acquired by the distributed antenna array, the explosion free energy is predicted using a generative adversarial digital twin model. When the current or predicted explosion free energy exceeds a threshold, a strategy rollback command or an emergency shutdown command is sent to achieve closed-loop control.

2. The method according to claim 1, characterized in that, The negative pressure value is obtained by an absolute differential pressure sensor, and the oxygen volume fraction is obtained by a zirconium-based oxygen sensor. The negative pressure value, oxygen volume fraction, and temperature-strain information are aggregated with a unified timestamp to form environmental data.

3. The method according to claim 1, characterized in that, Before injection, the inert gas mixture is continuously mixed with carbon dioxide and nitrogen at a preset volume fraction by a mass flow control device. The mixing flow rate is adjusted in real time according to the oxygen volume fraction to keep the oxygen volume fraction below the safe threshold.

4. The method according to claim 1, characterized in that, The microphone array is arranged at a non-uniform angle along the outer wall of the adjustable flow angle cyclone separator, and the sound pressure data acquired by the microphone array is sparsely reconstructed using the compressed sensing method to obtain the sound node amplitude.

5. The method according to claim 1, characterized in that, The adjustable guide angle cyclone separator synchronously adjusts the guide angle and cone ratio through an electric actuator. The guide angle and cone ratio are adjusted in a closed loop according to the real-time changes in the acoustic node amplitude and the outlet dust concentration.

6. The method according to claim 1, characterized in that, The dielectric barrier discharge plasma reactor maintains a stable ionization state and suppresses spark generation by adjusting the voltage and spacing between electrodes to match the discharge intensity with the oxygen volume fraction.

7. The method according to claim 1, characterized in that, Waste heat driven thermoacoustic transducers receive waste heat generated by dielectric barrier discharge plasma reactors and release acoustic energy at a fixed acoustic frequency to enhance the coagulation of fine dust particles.

8. The method according to claim 1, characterized in that, The microthermochemical model calculates the explosion free energy and system entropy and outputs the gradient. The reinforcement learning network updates the control strategy based on the gradient to generate control commands for adjusting the negative pressure pulse, inert gas flow rate, sound field intensity and cyclone geometry parameters.

9. The method according to claim 8, characterized in that, The reinforcement learning network includes a graph neural network to describe the coupling relationships between devices in the dust removal system, and a time encoder to process the time-series features of environmental data and captured data.

10. The method according to claim 1, characterized in that, The distributed antenna array is deployed in the form of a planar array. It obtains radio frequency reflection amplitude and phase information by scanning the working area through phase modulation to form radio frequency imaging data. It generates an adversarial digital twin model, which combines fluid dynamics calculation and discrete element calculation and perturbs the Green tensor to simulate the most unfavorable working conditions. When the current or predicted explosion free energy exceeds the threshold, the emergency shutdown command raises the negative pressure pulse and inert gas flow rate to the preset maximum safe value, and keeps the sound field intensity and the guide angle and cone ratio of the adjustable guide angle cyclone separator at the most recent safe state.

Images